Cover Image

SafeWork-R1: Coevolving Safety and Intelligence under the AI-45$^{\circ}$ Law

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Lab, Shanghai AI, :, Bao, Yicheng, Chen, Guanxu, Chen, Mingkang, Chen, Yunhao, Chen, Chiyu, Chen, Lingjie, Chen, Sirui, Chen, Xinquan, Cheng, Jie, Cheng, Yu, Deng, Dengke, Ding, Yizhuo, Ding, Dan, Ding, Xiaoshan, Ding, Yi, Dong, Zhichen, Du, Lingxiao, Fan, Yuyu, Feng, Xinshun, Fu, Yanwei, Gao, Yuxuan, Ge, Ruijun, Gu, Tianle, Gui, Lujun, Guo, Jiaxuan, He, Qianxi, Hou, Yuenan, Hu, Xuhao, Huang, Hong, Huang, Kaichen, Huang, Shiyang, Jiang, Yuxian, Lei, Shanzhe, Li, Jie, Li, Lijun, Li, Hao, Li, Juncheng, Li, Xiangtian, Li, Yafu, Li, Lingyu, Li, Xueyan, Liang, Haotian, Liu, Dongrui, Liu, Qihua, Liu, Zhixuan, Liu, Bangwei, Liu, Huacan, Liu, Yuexiao, Liu, Zongkai, Lu, Chaochao, Lu, Yudong, Lu, Xiaoya, Lu, Zhenghao, Lv, Qitan, Ma, Caoyuan, Ma, Jiachen, Ma, Xiaoya, Ma, Zhongtian, Meng, Lingyu, Miao, Ziqi, Niu, Yazhe, Peng, Yuezhang, Pu, Yuan, Qi, Han, Qian, Chen, Qiao, Xingge, Qu, Jingjing, Qu, Jiashu, Qu, Wanying, Qu, Wenwen, Qu, Xiaoye, Ren, Qihan, Ren, Qingnan, Ren, Qingyu, Shao, Jing, Shao, Wenqi, Shao, Shuai, Shi, Dongxing, Song, Xin, Song, Xinhao, Teng, Yan, Tong, Xuan, Wang, Yingchun, Wang, Xuhong, Wang, Shujie, Wang, Xin, Wang, Yige, Wang, Yixu, Wang, Yuanfu, Wang, Futing, Wang, Ruofan, Wang, Wenjie, Wang, Yajie, Wei, Muhao, Wen, Xiaoyu, Weng, Fenghua, Wu, Yuqi, Xiong, Yingtong, Xu, Xingcheng, Yang, Chao, Yang, Yue, Yao, Yang, Ye, Yulei, Yin, Zhenyun, Yu, Yi, Zhang, Bo, Zhang, Qiaosheng, Zhang, Jinxuan, Zhang, Yexin, Zheng, Yinqiang, Zhou, Hefeng, Zhou, Zhanhui, Zhu, Pengyu, Zhu, Qingzi, Zhu, Yubo, Zhou, Bowen
Format:	Preprint
Published:	2025
Subjects:	Artificial Intelligence Computation and Language Computer Vision and Pattern Recognition
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

_version_	1866916884700463104
author	Lab, Shanghai AI : Bao, Yicheng Chen, Guanxu Chen, Mingkang Chen, Yunhao Chen, Chiyu Chen, Lingjie Chen, Sirui Chen, Xinquan Cheng, Jie Cheng, Yu Deng, Dengke Ding, Yizhuo Ding, Dan Ding, Xiaoshan Ding, Yi Dong, Zhichen Du, Lingxiao Fan, Yuyu Feng, Xinshun Fu, Yanwei Gao, Yuxuan Ge, Ruijun Gu, Tianle Gui, Lujun Guo, Jiaxuan He, Qianxi Hou, Yuenan Hu, Xuhao Huang, Hong Huang, Kaichen Huang, Shiyang Jiang, Yuxian Lei, Shanzhe Li, Jie Li, Lijun Li, Hao Li, Juncheng Li, Xiangtian Li, Yafu Li, Lingyu Li, Xueyan Liang, Haotian Liu, Dongrui Liu, Qihua Liu, Zhixuan Liu, Bangwei Liu, Huacan Liu, Yuexiao Liu, Zongkai Lu, Chaochao Lu, Yudong Lu, Xiaoya Lu, Zhenghao Lv, Qitan Ma, Caoyuan Ma, Jiachen Ma, Xiaoya Ma, Zhongtian Meng, Lingyu Miao, Ziqi Niu, Yazhe Peng, Yuezhang Pu, Yuan Qi, Han Qian, Chen Qiao, Xingge Qu, Jingjing Qu, Jiashu Qu, Wanying Qu, Wenwen Qu, Xiaoye Ren, Qihan Ren, Qingnan Ren, Qingyu Shao, Jing Shao, Wenqi Shao, Shuai Shi, Dongxing Song, Xin Song, Xinhao Teng, Yan Tong, Xuan Wang, Yingchun Wang, Xuhong Wang, Shujie Wang, Xin Wang, Yige Wang, Yixu Wang, Yuanfu Wang, Futing Wang, Ruofan Wang, Wenjie Wang, Yajie Wei, Muhao Wen, Xiaoyu Weng, Fenghua Wu, Yuqi Xiong, Yingtong Xu, Xingcheng Yang, Chao Yang, Yue Yao, Yang Ye, Yulei Yin, Zhenyun Yu, Yi Zhang, Bo Zhang, Qiaosheng Zhang, Jinxuan Zhang, Yexin Zheng, Yinqiang Zhou, Hefeng Zhou, Zhanhui Zhu, Pengyu Zhu, Qingzi Zhu, Yubo Zhou, Bowen
author_facet	Lab, Shanghai AI : Bao, Yicheng Chen, Guanxu Chen, Mingkang Chen, Yunhao Chen, Chiyu Chen, Lingjie Chen, Sirui Chen, Xinquan Cheng, Jie Cheng, Yu Deng, Dengke Ding, Yizhuo Ding, Dan Ding, Xiaoshan Ding, Yi Dong, Zhichen Du, Lingxiao Fan, Yuyu Feng, Xinshun Fu, Yanwei Gao, Yuxuan Ge, Ruijun Gu, Tianle Gui, Lujun Guo, Jiaxuan He, Qianxi Hou, Yuenan Hu, Xuhao Huang, Hong Huang, Kaichen Huang, Shiyang Jiang, Yuxian Lei, Shanzhe Li, Jie Li, Lijun Li, Hao Li, Juncheng Li, Xiangtian Li, Yafu Li, Lingyu Li, Xueyan Liang, Haotian Liu, Dongrui Liu, Qihua Liu, Zhixuan Liu, Bangwei Liu, Huacan Liu, Yuexiao Liu, Zongkai Lu, Chaochao Lu, Yudong Lu, Xiaoya Lu, Zhenghao Lv, Qitan Ma, Caoyuan Ma, Jiachen Ma, Xiaoya Ma, Zhongtian Meng, Lingyu Miao, Ziqi Niu, Yazhe Peng, Yuezhang Pu, Yuan Qi, Han Qian, Chen Qiao, Xingge Qu, Jingjing Qu, Jiashu Qu, Wanying Qu, Wenwen Qu, Xiaoye Ren, Qihan Ren, Qingnan Ren, Qingyu Shao, Jing Shao, Wenqi Shao, Shuai Shi, Dongxing Song, Xin Song, Xinhao Teng, Yan Tong, Xuan Wang, Yingchun Wang, Xuhong Wang, Shujie Wang, Xin Wang, Yige Wang, Yixu Wang, Yuanfu Wang, Futing Wang, Ruofan Wang, Wenjie Wang, Yajie Wei, Muhao Wen, Xiaoyu Weng, Fenghua Wu, Yuqi Xiong, Yingtong Xu, Xingcheng Yang, Chao Yang, Yue Yao, Yang Ye, Yulei Yin, Zhenyun Yu, Yi Zhang, Bo Zhang, Qiaosheng Zhang, Jinxuan Zhang, Yexin Zheng, Yinqiang Zhou, Hefeng Zhou, Zhanhui Zhu, Pengyu Zhu, Qingzi Zhu, Yubo Zhou, Bowen
contents	We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety. It is developed by our proposed SafeLadder framework, which incorporates large-scale, progressive, safety-oriented reinforcement learning post-training, supported by a suite of multi-principled verifiers. Unlike previous alignment methods such as RLHF that simply learn human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, giving rise to safety `aha' moments. Notably, SafeWork-R1 achieves an average improvement of $46.54\%$ over its base model Qwen2.5-VL-72B on safety-related benchmarks without compromising general capabilities, and delivers state-of-the-art safety performance compared to leading proprietary models such as GPT-4.1 and Claude Opus 4. To further bolster its reliability, we implement two distinct inference-time intervention methods and a deliberative search mechanism, enforcing step-level verification. Finally, we further develop SafeWork-R1-InternVL3-78B, SafeWork-R1-DeepSeek-70B, and SafeWork-R1-Qwen2.5VL-7B. All resulting models demonstrate that safety and capability can co-evolve synergistically, highlighting the generalizability of our framework in building robust, reliable, and trustworthy general-purpose AI.
format	Preprint
id	arxiv_https___arxiv_org_abs_2507_18576
institution	arXiv
publishDate	2025
record_format	arxiv
spellingShingle	SafeWork-R1: Coevolving Safety and Intelligence under the AI-45$^{\circ}$ Law Lab, Shanghai AI : Bao, Yicheng Chen, Guanxu Chen, Mingkang Chen, Yunhao Chen, Chiyu Chen, Lingjie Chen, Sirui Chen, Xinquan Cheng, Jie Cheng, Yu Deng, Dengke Ding, Yizhuo Ding, Dan Ding, Xiaoshan Ding, Yi Dong, Zhichen Du, Lingxiao Fan, Yuyu Feng, Xinshun Fu, Yanwei Gao, Yuxuan Ge, Ruijun Gu, Tianle Gui, Lujun Guo, Jiaxuan He, Qianxi Hou, Yuenan Hu, Xuhao Huang, Hong Huang, Kaichen Huang, Shiyang Jiang, Yuxian Lei, Shanzhe Li, Jie Li, Lijun Li, Hao Li, Juncheng Li, Xiangtian Li, Yafu Li, Lingyu Li, Xueyan Liang, Haotian Liu, Dongrui Liu, Qihua Liu, Zhixuan Liu, Bangwei Liu, Huacan Liu, Yuexiao Liu, Zongkai Lu, Chaochao Lu, Yudong Lu, Xiaoya Lu, Zhenghao Lv, Qitan Ma, Caoyuan Ma, Jiachen Ma, Xiaoya Ma, Zhongtian Meng, Lingyu Miao, Ziqi Niu, Yazhe Peng, Yuezhang Pu, Yuan Qi, Han Qian, Chen Qiao, Xingge Qu, Jingjing Qu, Jiashu Qu, Wanying Qu, Wenwen Qu, Xiaoye Ren, Qihan Ren, Qingnan Ren, Qingyu Shao, Jing Shao, Wenqi Shao, Shuai Shi, Dongxing Song, Xin Song, Xinhao Teng, Yan Tong, Xuan Wang, Yingchun Wang, Xuhong Wang, Shujie Wang, Xin Wang, Yige Wang, Yixu Wang, Yuanfu Wang, Futing Wang, Ruofan Wang, Wenjie Wang, Yajie Wei, Muhao Wen, Xiaoyu Weng, Fenghua Wu, Yuqi Xiong, Yingtong Xu, Xingcheng Yang, Chao Yang, Yue Yao, Yang Ye, Yulei Yin, Zhenyun Yu, Yi Zhang, Bo Zhang, Qiaosheng Zhang, Jinxuan Zhang, Yexin Zheng, Yinqiang Zhou, Hefeng Zhou, Zhanhui Zhu, Pengyu Zhu, Qingzi Zhu, Yubo Zhou, Bowen Artificial Intelligence Computation and Language Computer Vision and Pattern Recognition We introduce SafeWork-R1, a cutting-edge multimodal reasoning model that demonstrates the coevolution of capabilities and safety. It is developed by our proposed SafeLadder framework, which incorporates large-scale, progressive, safety-oriented reinforcement learning post-training, supported by a suite of multi-principled verifiers. Unlike previous alignment methods such as RLHF that simply learn human preferences, SafeLadder enables SafeWork-R1 to develop intrinsic safety reasoning and self-reflection abilities, giving rise to safety `aha' moments. Notably, SafeWork-R1 achieves an average improvement of $46.54\%$ over its base model Qwen2.5-VL-72B on safety-related benchmarks without compromising general capabilities, and delivers state-of-the-art safety performance compared to leading proprietary models such as GPT-4.1 and Claude Opus 4. To further bolster its reliability, we implement two distinct inference-time intervention methods and a deliberative search mechanism, enforcing step-level verification. Finally, we further develop SafeWork-R1-InternVL3-78B, SafeWork-R1-DeepSeek-70B, and SafeWork-R1-Qwen2.5VL-7B. All resulting models demonstrate that safety and capability can co-evolve synergistically, highlighting the generalizability of our framework in building robust, reliable, and trustworthy general-purpose AI.
title	SafeWork-R1: Coevolving Safety and Intelligence under the AI-45$^{\circ}$ Law
topic	Artificial Intelligence Computation and Language Computer Vision and Pattern Recognition
url	https://arxiv.org/abs/2507.18576